Google Research and Google DeepMind tracked “Team Agent,” a persistent AI deployed across more than 20 internal teams, for five months and 41,000 conversational turns. The behaviors sold as the agent’s value — joining conversations unprompted, DMing people, surfacing documents — turned out to be exactly what broke participants’ trust, because the agent was making calls about unwritten social rules it had no way to know.
Two hosts work through the study’s sharpest quotes, the same-day launch of Ando, a Slack rival that sells the very unprompted behaviors the study flags, and the study’s own polarized data: a companion editor study showing timing and consent, not proactivity itself, decide whether an intervention lands or gets dismissed.